Skip to content

quickstart.md

Latest commit

 

History

History
105 lines (78 loc) · 2.61 KB

quickstart.md

File metadata and controls

105 lines (78 loc) · 2.61 KB

Quickstart

Prerequisites

  • Python 3.10+
  • An OpenCode server (or the opencode CLI installed locally)
  • An API key for your LLM provider (e.g., ANTHROPIC_API_KEY)

Install

npx skills add signalrush/auto

Or manually:

pip install auto-agent

Start OpenCode

# Install opencode
npm install -g opencode-ai

# Set your API key
export ANTHROPIC_API_KEY=sk-ant-...

# Start the server
opencode serve --port 54321

Write a program

# program.py
async def main(auto):
    # A single remind — one agent turn
    result = await auto.remind("What is 2 + 2? Just the number.")
    print(result)  # "4"

    # A remind with tool use
    result = await auto.remind("List all Python files in the current directory.")
    print(result)

    # A remind with structured output
    result = await auto.remind(
        "Count the lines in each Python file.",
        schema={"files": [{"name": "str", "lines": "int"}]}
    )
    print(result)  # {"files": [{"name": "main.py", "lines": 42}, ...]}

Run it

auto-run program.py

Write a loop

The power of remind() is that the model controls the work, but Python controls the state:

# autoresearch.py
async def main(auto):
    # Get baseline
    baseline = await auto.remind(
        "Run `python train.py` and report the validation loss.",
        schema={"val_loss": "float"}
    )
    best = baseline["val_loss"]

    for i in range(20):
        result = await auto.remind(
            "Propose and run one experiment. Edit train.py, run it, report results.",
            schema={"val_loss": "float", "description": "str"}
        )

        if result["val_loss"] < best:
            best = result["val_loss"]
            print(f"New best: {best} — {result['description']}")
        else:
            await auto.remind("Revert the last change. Git reset to previous commit.")

Run with auto-run autoresearch.py. The session persists across all steps — the model remembers every experiment. The Python loop handles state, branching, and control flow. The model handles the actual coding work inside each remind.

Multi-agent

async def main(auto):
    auto.agent("tester", cwd="/home/user/project")

    await auto.remind("implement the new feature")
    test_result = await auto.task("run the full test suite and report", to="tester")
    await auto.remind(f"Fix issues from tests: {test_result}")

Next steps

  • API Reference — full auto.remind(), auto.task(), and auto.agent() docs
  • Design — why auto works this way
  • Examples — complete example programs